All the Patents aims to improve the patent system by using AI to generate and publish “obvious” inventions, making them searchable for patent examiners and the public to prevent unwarranted patent monopolies.

Technical Approach

kl3m-170 model: Small GPT2-sized language model trained on copyright-free data
Low resource requirements: Runs on ~1GB VRAM, CPU, or Apple Silicon
Two generation methods:
- Top-down: Creating detailed claims from abstract patent information
- Bottom-up: Creating abstracts and titles from detailed claims

Core Philosophy

The project poses a fundamental question: “If a GPT2-sized model can generate a valid set of claims, should anyone be able to monopolize the invention?”

Tools and Implementation

Streamlit application for model download and generation
API and client scripts for programmatic access
VLLM integration for efficient inference
Research focus on data collection, model training, and hit rate evaluation

Impact Goals

Address patent system barriers that stifle innovation
Make obvious inventions searchable and accessible
Foster genuine innovation by preventing trivial patent monopolies
Demonstrate AI’s capability to generate patent-like claims

Current Status

Early-stage research and development focusing on prototype tools, generation techniques, and evaluation metrics for generated patent claims.